Presentations
-
Hyperparameter Transfer in Graph Neural Networks
A parameterization for GNNs under which a single base learning rate, tuned once on a small model, carries over as width and depth grow. Joint work with Boris Hanin.
-
Principles for Initialization and Architecture Selection in Graph Neural Networks
Three ways a deep GNN can fail before training starts — exponential outputs, oversmoothing, and correlation collapse — and the initialization and architecture choices that fix each one. Joint work with Boris Hanin.
-
Hyperparameter Transfer: A Recipe for Efficient and Robust Scaling
Re-tuning large models is expensive. This talk lays out a recipe for tuning small GNNs and transferring their hyperparameters to much larger ones, with stable training dynamics across scale.